Pointing With Concepts: Explorations in the Embodiment of Knowledge

Rajesh Kasturirangan
1. Introduction. Common sense avers that when you tell me something (say, that the
earth moves around the sun) and you are not lying, then I come to know that something
(that the earth moves around the sun, for example). What is the nature of this knowledge?
Where does it come from and where does it reside? If one turned to western philosophy
to answer this question, one would be hard pressed to find an answer to this question,
primarily because western philosophers starting with Plato have been loath to accept
testimony as a source of knowledge. Furthermore, this reluctance to accept testimony has
led to one of the most intractable debates in the western epistemology, about whether
knowledge is innate or learnt. Heres what Socrates had to say on this topic (in Meno):
Fragment 1
SOCRATES: Attend now to the questions which I ask him, and observe whether he
learns of me or only remembers.
MENO: I will.
Fragment 2
SOCRATES: And you know that a square figure has these four lines equal?
BOY: Certainly.
SOCRATES: A square may be of any size?
BOY: Certainly.
Fragment 3
SOCRATES: And does not this line, reaching from corner to corner, bisect each of these
spaces?
BOY: Yes.
SOCRATES: And are there not here four equal lines which contain this space?
BOY: There are.
1
Figure 1
Fragment 4
SOCRATES: That is, from the line which extends from corner to corner of the figure of
four feet?
BOY: Yes.
SOCRATES: And that is the line which the learned call the diagonal. And if this is the
proper name, then you, Meno's slave, are prepared to affirm that the double space is the
square of the diagonal?
BOY: Certainly, Socrates.
SOCRATES: What do you say of him, Meno? Were not all these answers given out of
his own head?
MENO: Yes, they were all his own.
SOCRATES: And yet, as we were just now saying, he did not know?
MENO: True.
SOCRATES: But still he had in him those notions of his--had he not?
MENO: Yes.
SOCRATES: Then he who does not know may still have true notions of that which he
does not know?
MENO: He has.
Platos arguments have been enormously influential, affecting cognitive science to this
day, but they leave us unclear about some fundamental questions. What information was
transmitted from Socrates to the slave boy? What did the slave boy learn
1
? It is hard to
see how an information theoretic model can help us understand these questions. On the
other hand, Platos arguments remain convincing even now
2
. How else can we explain
the fact that we know that squares have equal sides or that they can have any size
given that real world objects have neither of those two properties?
The basic argument is based on the intuition that correct responses to questions are
predicated on intrinsic knowledge. Is that intuition correct? It is not clear to me if thats
the case. For example, if you show me two lines placed side by side and ask me which
one is longer, I will answer immediately, even though I didnt know before hand which
one was longer all I had to do was to look at the world and do the comparison in the
world. In other words, the world can help me answer the question without me having to
store all the answers in my head beforehand. For abstract concepts though, one could say
1
Actually, Plato answered these questions as follows: (a) Socrates did not transmit anything; and (b) the
slave boy learnt nothing, but for those of us who persist in thinking that something was taught and
something was learnt, we would like to know what is available to be known from testimony. To use a
Gibsonian term, what are the affordances of the spoken word?
2
In this talk, I am not going to mention the classical Indian debates on the nature of knowledge in
testimony. Not that the Indian material is unimportant, but that we need to work through the western
material to see where it goes, and secondly, once we see the points where the western arguments break
down, we would have a more natural interface with any insights that one might bring in from the Indian
traditions.
2
that the appropriate stimulus is not available in the world at all. A Platonist might argue,
for example, that nothing in the world comes pre-labeled as beautiful, so one couldnt
learn the meaning of the concept beauty by staring at the world. They would say that
the mind imposes beauty upon the world, which has no aesthetics of its own.
If so, then how does the mind know what to apply those concepts to? Is it entirely
arbitrary? Clearly not, since we find symmetric objects more beautiful than asymmetric
ones. What is it about symmetry that makes an object beautiful? The problem with innate
accounts of concepts is that while they explain the abstract features of concepts, they
dont tell us anything about how those concepts help us know their objects.
2. Theories of Concepts. I started talking about knowledge but then I shifted my focus to
concepts, which I know is too fast a shift. There is a long history of the importance of
concepts for epistemology and the philosophy of mind, but I will not entertain you with
that history today. Instead I am going to ward off your queries by telling you two reasons
to think about concepts in relation to knowledge. I should say at the outset that my
approach is going to sound very reductive to those of you who are used to thinking about
concepts and conceptual change in terms of history and social forces etc. I will be talking
about very mundane concepts like IN and ON, which might seem very far removed from
notions like freedom or truth. Which would be fine, but the implicit claim I am making is
that knowing these bread and butter concepts will help us understand the general
processes of conceptual change in society as a whole. I am not going to defend that thesis
today, but I will talk about it at the end of the lecture.
(1) Compositionality. If we think of knowledge as the grasping of true propositions, and
propositions are made of concepts and their truth is derivative of the truth of particular
concepts, then we need to think about concepts in order to understand knowledge.
(2) Indirect knowing. Since we dont know the world directly and only through our
representations and concepts are the essence of our representations, we need to look at
concepts.
Therefore, without much ado, I am going to shift to the study of concepts. Let us start
with some current accounts of concepts. Here I outline two broad theories of concepts
and their problems.
2A. Concepts as Definitions. Here we assume that the meaning of a concept is given by
some necessary and sufficient conditions. For example, we could define the concept
BACHELOR as unmarried man. Unfortunately, outside mathematics and certain areas
of science, definitions dont work (not surprisingly, the definitional account of concepts
was based on their success in mathematics). Is the Pope a bachelor? Perhaps we could
solve the problem by redefining BACHELOR as unmarried man who is not celibate of
his own free will. Then what about the Medieval Popes, like the Borgias, who were
certainly not celibates? Are they Bachelors? The definitional account runs quickly into
three insuperable problems:
(1) As I said above, necessary and sufficient conditions are hard to find.
3
(2) Concepts have prototypical and not so prototypical examples, while definitions do
not distinguish between one exemplar of a concept and another. For example, a
penguin seems less prototypical of the concept BIRD than a sparrow. Why?
(3) Definitions use concepts that themselves require definitions, which leads us to
Fodors thesis that all concepts are innate, including ones such as AIRPLANE.
While the definition account is certainly biased towards a rationalist theory of concepts
(since they necessary and sufficient criteria are all in my head), a parallel empiricist
theory is not that hard to find, which is to identify the meaning of a concept with all the
things in the world it applies to, so for example the concept DOG would mean all the
dogs in the world. Here too one would run into difficulties, for example, how does one
learn abstract things about purely empirical concepts? How does one know that a dog has
a certain DNA if one is only exposed to visual instances of dogs? The same critique goes
for feature based accounts of concepts. If by DOG, one means a list of features such as:
FOUR LEGGED, ANIMAL, MAMMAL, BARKS etc, then how does one account for
three legged dogs? And what about the fact the analytic entailment: DOG MAMMAL.
Doesnt that follow from the meaning of the concept DOG? How do features account for
such entailments?
2B. Concepts as Prototypes. The theory of concepts as based on definitions was driven
by the desire to account for concepts as the building blocks of propositions. However, we
can also think of concepts as being the basis for categorization. For example, an early
hominid would have needed to classify an animal as predator or prey very quickly. A
snap categorization is in order to tell the motor system to hunt or to flee. Prototypes
allow us to classify a stimulus quickly and to convey the results of that classification to
the motor system. For doing so, we may need multiple prototypes for the same category.
For example, the category BIRD may have a sparrow or an eagle as the normal prototype
but penguins as prototypes for flightless bird. Research by Eleanor Rosch and others has
showed that categories are organized hierarchically (EAGLE-BIRD-ANIMAL-LIVING
THING, for example) and that furthermore, there is a basic level of categorization where
processing and categorization happens automatically and quickly, faster than at any other
level (for most people that would be EAGLE).
Note that the prototype theory handles graded categories quite easily, since membership
in a category is proportional to the similarity to a given prototype. However, prototype
theory doesnt explain vagueness: for example, when does a man become tall? Its not just
that we have different prototypes for tall people in different cultures but that in a given
culture, we cannot even draw a boundary. If grading was valid then, there would be a
category boundary (at the very least a fuzzy boundary), but it seems that there is none to
be drawn. Vagueness is a problem for the prototype from the fact that there seems to be
no basis for the transition from one category to another, while the prototype theory where
postulate one. Furthermore, how do I know which prototype to use? Why should I use
sparrows when I see birds? Doesnt that imply some knowledge of birds to start with?
Where is that knowledge accounted for in the prototype account?
4
There is a development in the prototype account that one needs to note, which is the
metaphor theory of concepts. Embodiment theorists will say that prototypes are really
sensorimotor clusters that connect stimuli to categories. Furthermore, they say that even
for abstract concepts like TIME or BEAUTY, we use particular sensorimotor clusters as
metaphorical prototypes, like when I say Time is like a river. A particular image of a
river may be what I use to understand the concept TIME as a whole. Note that here
prototypes are not being used as exemplars for a category, but using the exemplars of one
category to structure another. Metaphor theory has all the advantages of the prototype
theory and it also gives an explanation for how we learn abstract concepts, and hence to
Platos problem. But metaphor theory doesnt have an answer for vagueness, and it does
not have an answer to the question why does a metaphor give us knowledge? Of course
Plato would have said that knowledge was not going to happen anyway and therefore
listen to the metaphor theorist, but those of us who disagree with Plato cannot be satisfied
with a theory of concepts based on metaphor.
Now suppose we want to keep the embodied account of concepts but are not willing to
settle for a non-knowing account of concepts. Then what can we do? In the next section, I
develop a new account of concepts based on knowing concepts as a knowing how
rather than a knowing that.
2C. Pointing with Concepts. All of these theories of concepts model knowing a concept
as knowing that rather than a knowing how. Knowing that is an abstract, theoretical
kind of knowing, for example when I know that the earth goes around the sun, while
knowing how is knowledge based on action, knowing how to ride a bicycle. Except that
we dont know what kind of knowledge it is to know a concept. Is it knowledge of
meaning or knowledge of the world? The two models that I have discussed so far do not
admit any knowledge of the world as part of the knowledge of these concepts. But in
what way do I get to know anything about time when I say time is like a river? Do I
just get to know what usage is legitimate, or do I actually know something about time?
Metaphor theorists would like us to believe the second, but they dont tell us how we
acquire knowledge about time from these metaphors. There is no notion of judgment or
veridicality built into these two notions of concepts. Is there an alternative? These days,
there is a theory of knowledge which says that perception, for example, is enactive,
which would make it a knowing how rather than a knowing that. Traditionally, seeing an
object is modeled as constructing a three dimensional model of the object in our minds,
so that we know how the object looks like from all direction. In the enactive approach,
we dont try to build detailed representations at all. For the most part, we dont need
those details, so why bother. Instead we rely on the fact that all the details of the object
are available in principle, by moving our bodies around the object and saccading our eyes
to the spots where we need to see more detail. That cup thats in the other room, I dont
need to know what it looks like from here. All I need to know is that if I walk next door, I
will see the cup sitting on the table and if I need to see the intricate etching on the cups
surface, I can pick it up and bring it closer to my eyes. Why should I have to represent all
this detail? Representation is expensive, and the world is always there to help me out
5
when I need it. In this paper I want to extend this enactive account to the meaning of
concepts.
For a moment, imagine that the meaning of a concept was analogous to an object of
perception. Suppose we encounter an object while moving about in this world. The object
presents itself to you with a certain face, the front of the object from where you are
standing. The back of the object is obscured, but you know its there you might even
have an inkling as to what the back looks like, as in the figure below where you might be
predisposed to believe that the object is a square.
Figure 1
Now, what if the meaning of a concept was like seeing an object? It would imply that
when we encounter the meaning of a concept, we would always encounter it from a
certain view. Sometimes the point of view is conventionalized, so that we have a default
meaning for that concept though there can be different default meanings depending on
the context. For example, when I say to you I am standing in front of the museum, you
understand that I am located near the side of the building that is conventionally the front
of the building (note that the front of a building can depend on whether you are inside or
outside you enter the front of the church and suddenly you are at the back of it!). On the
other hand, if I say to you he is standing in front of the tree, I mean that he standing
between my line of sight and the tree, since the tree does not have a front. In other words,
the meaning of a concept like IN FRONT OF can have different views depending on the
context.
Furthermore, this perceptual analogy also explains why concepts are naturally vague
for we dont know the backside of concepts (or objects) in anything like the way the way
we know their fronts. The backside of an object is an absence thats present to our
perceptual apparatus. The same would go for the meaning of concepts - that we wouldnt
know what its hidden layers of meaning look like even though we know that there are
some. Therefore, depending on the view, we wouldnt want to commit ourselves to what
an object or concepts looks like in the back just by looking at the front. In other words,
the meaning of a concept is always open to other possibilities depending on your view of
it, even if you have a strong bias to thinking about it in one way, just as the object in the
previous figure is likely to be a square but could be something else. Vagueness, in this
understanding of concepts, is a natural consequence of the fact that the meaning of a
concept is not entirely visible in one view. Furthermore, the metaphorical structuring of
concepts becomes a natural consequence of exploring the full object thats constitutes
the meaning of the object.
6
So far, so good. But can one make this insight into concepts as objects into a precise
theory? Can one have a mathematical theory of concepts based on this intuition? Heres
where we will run into some deep conceptual problems related to the nature of the object
that the meaning of a concept could possibly be, problems related to the fact that while
the representations of perceptual objects are underdetermined by their interface with our
perceptual systems, the representations of concepts are infinitely undetermined. So the
technical problems go much deeper. And then, we have to eventually face the issue of the
origin of conceptual objects squarely in the face while it was possible for Gibson, for
example, to say that the objects of perception are out there in the world, its not even clear
what world contains the meaning of conceptual objects. All of these philosophical topics
will be addressed in the later sections, but for now, we need to get down to some real
computational details. In the next section, we will see how to flesh out the insight that the
meaning of concepts is like a perceptual object and then apply that theory to a model
class of concepts, namely, prepositions.

3 A Theory of prepositional semantics.
The semantics of prepositions can be understood in terms of derivations where the notion
of derivation was defined in chapter 3. In the case of prepositions, I propose the
following semantic hypothesis:
Hypothesis: Let
1
,
2
,
3
,
4
be categories (in the sense of chapter 2) and let P
1
,P
2
,
P
3
, P
4
be the partial orders generated by these categories. Let C
1
,C
2
, C
3
, C
4
be
concepts, i.e., nodes in the respective partial orders. The meaning of a preposition is a
four step derivation of concepts C
1
, C
2
, C
3
, C
4
where:
(a) C
1

1
is the topological, core meaning of the preposition.
(b) C
2

2
is an abstract augmentation of C
1
.
(c) C
3

3
is a concrete referent of C
2
.
(d) C
1

4
is a concrete augmentation of C
3
.
In the rest of this chapter, I will flesh out the derivation theory of prepositional
meaning. Before I go on to a detailed exposition of the derivation model of
prepositional meaning, let me make a few general remarks about the semantics of
prepositions.
3.1. Generating the meaning of prepositions.
The category theoretic approach to generating meanings depends on the existence
of global models that are relevant to a particular domain. These models encode the
representations that are functionally relevant to the tasks being performed by an agent
who is embedded in that domain. What is the functional importance of spatial
prepositions?
One important property of spatial prepositions is that they are locational, i.e., they
are used to communicate the location of a target object (often smaller and possibly
mobile) with respect to another reference object (often larger and static). We can see the
locational nature of spatial prepositions when we look possible answers to the following
question:
Question: Where is John?
7
Answers: (a) At school ; (b) In the house; (c) On the roof; (d) Across the park
In each of these cases, the location of John with respect to a reference object is
being communicated. On the other hand, it is hard to imagine a question using What or
Why that could get a plausible answer using a spatial preposition.
If spatial prepositions were purely locational, they could all be generated from
one global model, a model that encodes the locations of one object with respect to
another. Furthermore, since the location of an object in any coordinate frame is a purely
geometric relation, one would expect that their meaning is completely determined by the
geometrical relationship between the target and the reference objects. However, geometry
alone cannot capture the meaning of prepositions as we saw in figures 5.1 above, which
argues for the relevance of dynamic ideas of support. Does this mean that prepositions
are not locational?
Despite the counterexamples, there are several reasons why prepositions should still be
seen as being primarily locational. Let us look at some of them:
(1) The representation of the geometry of the figure and ground is asymmetric.
Prepositions do not care about the shape of the figure figure objects are
conceptualized as points (Talmy 2000). The shape of the ground object is
important prepositions are sensitive to topological features of the ground such as
its interior and boundary and to axial features of the ground as well. On the other
hand, the motion of the figure relative to the ground does matter while
prepositions do not let ground objects move with respect to figure objects at all.
This asymmetry makes sense only if we think of prepositions as encoding the
relative location of the figure with respect to the ground.
(2) Even simple non-locational relations are not encoded, even though they are as
simple as locational relations that are encoded. Consider the relations illustrated
in figure 5.3 along with a verb that describes the relation. HIT and BOUNCE are
not more complex than CONTAINMENT or CONTACT, so why are they not
encoded? One reason may be that HIT and BOUNCE are not locational.
(3) Even when non-locational relations are relevant, they are encoded only when
they are consistent with a locational role. For example, while support is
present in both figure 5.1 and figure 5.2, it is encoded only in figure 5.1,
which plausibly communicates the typical location of a book with respect to a
table.
8
Hit
Bounce
Figure 2
For these reasons, I believe that prepositions are fundamentally locational.
However, the notion of location is more complex than thought before. Both figure and
ground objects have degrees of freedom that need to be taken into account in a fully
developed model of location. For example, figure objects have a tendency to move
under gravity or on their own, if they are autonomous and also to change their state, say
from containment within the ground object to being outside the ground object.
In terms of the technical vocabulary of derivations, I argue that the core meaning ,
C
1
, (in my notation) is locational, but the augmented meanings, C
2
, C
3
andC
4
need not be
locational. In fact, the core meaning of prepositions has to be augmented to include
dynamical representations for both the figure object as well as the ground object. It is not
just the purely geometric relations that matter, geometry is a proper subset of dynamics.
One way to include dynamics is to include figural motion and change in figure state in
the figure models from the category
2
. Similarly, we can enrich the representation of the
ground objects in
2
by including physical characteristics of the ground object. Note that
even if the figure is not moving, the models needs to encode the fact that the lack of
motion is a dynamically maintained state (for example in figure 5.1 where the book and
the table are in a dynamically stable relationship that does not involve motion). The
category theory approach allows us to incorporate dynamic principles into the locational
models rather easily. Let us see how each one of these steps is worked out, starting with
the abstract, core meanings.
Step 1. Generating the abstract, core meanings of prepositions.
It seems reasonable to assume that spatial prepositions represent spatial relations
relevant to human beings acting in a dynamic spatio-temporal environment. Then it also
seems reasonable to assume that the abstract, core, concepts encoded by spatial
prepositions are those that belong to the most universally valid, i.e., highly invariant sub-
categories of our spatial environment. These common elements (borrowed from chapter
3, section 4) are:
(1) A gravitational environment with gravity pointing up.
(2) An autonomous agent (the perceiver), and possibly other agents.
(3) Solid objects of randomly distributed shapes, sizes and locations that
satisfy the following physical laws:
a. An object is at one place at one time
b. An object has fixed shape and size and mass.
c. An object moves in a continuous path.
(4) Most of the objects are motionless during the task, though some may
move due to forces exerted by an agent or by some other object.
(5) All of the agents and objects are located in an environment suffused with
light obeying the principles of geometric optics.
These constraints were modeled formally in chapter 3 as a model:
E = GP T
aff
O v
g
S
i
.
Where, E is the spatial environment, GP is the ground plane, are the T
aff
are the affine
transformations in 3-D space, O is the observer, v
g
is the vertical gravitational vector and
S
i
represent the smooth surfaces of solid objects in 3-D. The partial ordering generated by
9
the above model, E, has three broad sub-categories: Coordinate frames, Topological
Structure, and Metric Structure. A short outline of these categories is given below:
(1) Coordinate Frame Representations: In order of decreasing invariance, we
have three Coordinate frame representations- Gravitational Frame (GF), Blob-
location (BL), Principal Axis-location (PA) and Minor Axis-location (MA).
(2) Topological Structure Representations: Blobs (B), Dimension (D), Generic
spatial invariants (GI), Non-generic invariants, (NGI), Higher Topological
Invariants (HTI).
(3) Metric Structure Representations: Global Properties (GP), The Part Structure
Hierarchy (PSH), Convexity Hierarchy, (CH), Riemannian Structure (RS).
Metric Structure, Coordinate Frame and Topological Structure obey the following
partial ordering relations:
(a) Metric Structure Coordinate Frame
(b) Metric Structure Topological Structure
Furthermore, as shown earlier, when classified according to invariance,
Coordinate Frame, Topological Structure and Metric Structure can be decomposed into
the partially ordered categories given below:
(1) Coordinate Frame: MA
U
PA
U
BL
U
GF.
(2) Topological Structure: HTI
U
NGI
U
GI
U
D
U
B.
(3) Metric Structure: RS
U
CH, PSH
U
GP.
The partial ordering of E is the collection of all the partially ordered sequences given
above. Thus, a classification of spatial categories based on stability leads to a partially
ordered set of spatial categories.
The derivational hypothesis along with the inclusion relations in (a)-(b) and (1)-(3) above
make the following prediction:
The core meaning of prepositions must either be a coordinate frame
concept or a topological structure concept but not a metric structure
concept.
We can test this hypothesis rather easily. For simplicity, let us assume that each
preposition has a default, geometric meaning. For example, the default meaning of in is
containment which is a generic topological relation. A classification of geometric
features associated with the default meaning of prepositions, based on an (informal)
corpus analysis of The New York Times is given in the tables 1 and 2 below. While this
classification does not claim to be exhaustive, it serves to demonstrate that metric
features are not part of the meaning of prepositions. Even such seemingly metric
prepositions such as near are ordinal and not truly metric. Ordinal relations are
topological because they are invariant under all geometric transformations such as
dilations, and stretches.
Preposition About Above Across After Against Along Alongside Amid Among Around At
Central
Meaning
N/A GF GI GI NGI MA MA GI GI NGI BL
10
Preposition Forward Here Inward Left North Outward Right Sideways South There Together
Central
Meaning
MA BL GI MA MA GI MA MA MA BL NGI
Gravitational Frame, (GF), is the universal vertical frame defined by gravitation. Blob-location (BL),
represents a Blob at a particular location. Principal Axis-Location (PA), represents a principal axis of an
object at a given location while Minor Axis-Location (MA), represents a Minor Axis at a given location.
Blobs (B), are undifferentiated spatial representations, with no internal structure. Generic spatial invariants
(GI), are those spatial invariants that do not change when the stimulus is perturbed slightly. Non-generic
invariants, (NGI) are those spatial invariants that change when perturbed along certain direction.
Preposition Atop Behind Below Beneath Beside Between Betwixt Beyond By Down From
Central
Meaning
BL MA GF GF MA GI GI GI N/A GF BL
Preposition In Inside Into Near Nearby Off On Onto Opposite Out Outside
Central
Meaning
GI GI GI GI GI GI NGI NGI MA GI GI
Preposition Over Past Through Throughout To Toward Under Underneath Up Upon Via
Central
Meaning
GF GI GI GI BL BL MA MA GF BL BL
Preposition With Within Without Afterward Apart Away Back Backward Downstairs Downward East
Central
Meaning
N/A GI GI GI GI GI GI MA GF GF MA
Preposition Upstairs Upward West
Central
Meaning
GF GF MA
11
Table 1: Core meaning of Prepositions
Table 2: Prepositions by Spatial Type
Furthermore, note that prepositions do not name any topological or geometric
relations that have depth more than 2 in the partial ordering. For example, we do not have
a preposition that encodes the relationship between the rectangle and the black dot given
in figure 5.4 below.
We can say, The black dot is on the rectangle which does not capture the exact
geometrical relation or we can say The black dot as it the upper right corner of the
rectangle which is far more precise, but requires the qualifiers upper and right.
There is no single preposition that says at upper right corner. In other words, the core
meaning of a preposition only encodes an abstract spatial relation that has low
complexity, either 1 or 2.
Step 2: Augmenting Abstract Models
In the previous section, we saw that the core meanings of a preposition are
abstract locational predicates- either topological or coordinate frame predicates. In this
section, I will show how these core meanings can be augmented to incorporate dynamical
relations. Let FM stand for figure models and GM stand for ground models. Given the
earlier arguments for the locationality of prepositions, the articulation of figure and
ground models proceeds in different ways increasing dynamic capacities for the figure
and increasing geometric complexity for the ground.
Partial Order
Type
Gravitational
Frame
Generic invariant Blob Location Principal Axis OTHER
Frequency 10 30 9 16 4
12
Figure 3
The simplest model for both figure and ground is one where both figure and
ground are static points. Formally
FM(1) : Figure is a static point.
GM(1) Ground is a static point.
The next step is to add complexity to the figure by adding motion or change.
FM(2): Figure motion is modeled as point moving along a curve.
FM(3): Figure change is modeled as a change C: S U S where S and S are two
possible dynamically stable states. The simplest case of change, the one we will
use for prepositions is when there are only two states let say 0 and 1. In this case
FM(3) consists of four choices: 0 U 0, 0 U 1, 1 U 0, 1 U 1.
Similarly, one can augment the ground models as follows:
GM(2): Ground topology is articulated. The partial ordering of topological states
in the case of a 3D ground object is given below

GM(3): Ground axial structure is articulated. The partial ordering of axial
structure is given below:
Note that FM(1) is included (in the category theoretic sense as explained in
chapter 2) in FM(2) as well as FM(3). Similarly GM(1) is included in GM(2) as well as
GM(3). Now that we have these global models for the figure as well as the ground, we
can state the basic semantic hypothesis:
Prepositional Semantic Hypothesis: Suppose we map the figure models FM(1),
FM(2) and FM(3) on to the ground models GM(1), GM(2) and GM(3) using the
mapping principle described in chapter 3. Then, prepositional meanings are
generated by the partial orderings that can be derived from each one of these
mappings.
To see why the above hypothesis generates the appropriate meanings of
prepositions, let us take a look at the partial orders that are generated when the figure and
ground models are mapped on to each other and the prepositions that correspond to these
partial orders. Remember that according to the scheme developed in chapter 3, the partial
ordering comes from classifying the possible topological states of the combined figure-
ground models.
13
Exterior Interior Face Edge Vertex , U U U
MajorAxes MinorAxes U
Figure
Ground
AT NEXT
Ground Figure
NEAR
Ground Figure
BETWEEN
Ground Figure Ground
Figure 4(a). FM1 * GM1

14
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 5.5(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Ground
Figure
FROM
Ground
Figure
TO
Ground
Figure
AROUND
Figure 4(b). FM2 * GM1
Figure 4(c). FM1 * GM2
Figure
Ground
IN Figure
Ground
ON
FRONT
LEFT
RIGHT
BACK
Figure 4(d). FM1 * GM3
GROUND
Step 3. Mapping Abstract Meanings of Prepositions on to the Concrete.
In the previous section, I showed how the derivational approach can be used to
generate the correct meaning of prepositions under various circumstances. I argued that
the meaning of prepositions is determined by various models and that the semantics of
prepositions is determined by looking at the mappings between these models. In this
section, I show that the same approach can be used to understand the reference of
prepositions, i.e., what properties of the world are picked out by prepositions.
As I argued in the introduction and chapter 2, the problem of picking out the
worldly referent of a preposition is also a problem of mapping different models on to
15
Figure 4(e). FM3 * GM2
IN
GROUND
ON
GROUND
OUT
GROUND
OFF
GROUND
INTO
GROUND
ONTO
GROUND
OUT OF
GROUND
Figure 4(f). FM2 * GM3
ALONG
GROUND
each other. While the meaning of a preposition is abstract, its referent has to be a concrete
object in the world. How are these two mapped on to each other. My solution is to realize
that both the abstract and the concrete are representations, though these representations
make different aspects of the world explicit. Nevertheless, I will continue to assume that
the two representational schemes are generated by the same principle of partially ordered
categories described in chapter 3. In the previous section, I showed how the abstract
meaning of prepositions can be generated by a combination of topological, axial and
dynamic global models. In order to understand the mapping of prepositions on to the
world, we need to answer two questions:
(1) What is the generative model for the representation of concrete world
properties?
(2) Is the mapping from the abstract categories to the concrete categories given by the
mapping principles developed in chapter 3?
In the next few paragraphs, I show how one can answer these questions using
DERIVATIONAL framework. For simplicity, my discussion is restricted to the purely
locational, topological meaning of prepositions in a 2D Euclidean world. For locational,
topological meanings of prepositions in 2D, I show how one can model the concrete
world as a Metric category.
To be more formal, let us assume that the world consists of smooth curves and
their possible interiors in R
2
with the normal, Euclidean metric. One can then model this
world using two different categories, the topological and the metric.
A single curve in R
2
can only have one of two possible topological states - it can be either
open or closed. Similarly, given two smooth curves in 2D, their topological relations are
the following
(1) containment/outside,
(2) contact
(3) number of intersections
At the metric level, each curve can be modeled as follows:
A curve is a vector X(t) = | x(t), x/t , where x(t) is a parametrization of the
curve and x/t is its tangent vector field.
Then one can immediately define a few properties of the curve using the above
parametrization. In order to classify the metric category according to its partial ordering,
let us see how we can derive the properties of closure and intersection using the
parametrization developed above. For closure, note that
If a curve X(t) is closed, we can define its interior as follows:
The interior I(X) of a curve X(t) is the bounded component of R
2
X(t) (figure
5.6).
16

= closed is curve the if only and if integer an is

) ( 2
1
number winding The
t x
dx

Similarly, note that intersections can be defined as follows:

Two curves X(t) = | x(t), x/t , and Y(t) = | Y(t), y/t , intersect at a point t
0
if
and only if x(t
0
) = y(t
0
). Furthermore, the intersection is transversal if x/t|t=t0 U
y/t|t=t0 and it is non-transversal if x/t|t=t0 = y/t|t=t0.
In this way, we can see how there is a systematic way of mapping topological
categories to metric ones. Note that the mapping is an isomorphism of partial orders as
well the mapping is codimension preserving. For example, in the topological case, for
curves in R
2
, non-transversal intersections are codimension 1 intersections. Similarly,
when two curves in R
2
intersect, generically, the two tangent vector fields are not equal.
However, the set of possibilities when they are equal is of codimension 1. In other words,
the notion of non-transversality in the metric case has the same codimension as non-
transversality in the topological case. Therefore, the mapping from topological to metric
categories (shown above for the case of R
2
, but true more generally) is not just a mapping
of models, it is also a mapping of partial orders.
Step 4. Augmenting Concrete Models
The final piece in trying to understand the semantics of prepositions is to
understand how concrete models can be augmented. One may ask why this is necessary,
given that the semantics of prepositions is mainly abstract. The reason is because
concrete situations do make a difference to the meaning of a preposition. Consider the
two pairs of pictures in figure 5.7 below
17
Figure 6(a). Is the black egg in the bowl?
Exterior, E
1
Figure 6(b). Is the black book on the table?
Interior
Exterior
Figure 5
X(t)
In both figure 5.7(a) and 5.7(b), the figure and the ground have the same abstract
relationship, of containment in 5.7(a) and of support in 5.7(b). However they are judged
differently. The difference between the two cases cannot be attributed to the difference in
some abstractable feature. For example, one may think that support is the crucial feature
that is missing in 5.7(a) but it is present in 5.7(b) and yet the two cases are judged
differently. It is quite plausible that the semantic distinction between the two situations is
because of the metric and physical conditions that differentiate the two circumstances. In
other words, the meaning of a prepositions can reflect the concrete relations between the
figure, the ground and other objects in the scene.
I claim that we can understand the difference between the two situations by
looking at the way in which concrete spatial models can be augmented. As before, we
start with the hypothesis that spatial prepositions are primarily locational. However, as
we saw in the case of abstract models, the notion of location (in the abstract domain) has
to be augmented to include motion and change on part of the figure, leading to a richer
vocabulary for describing the semantics of prepositions. Therefore, it stands to reason
that the notion of location can be expanded in the concrete domain as well, and that this
expansion should lead to a richer vocabulary for describing prepositions in the concrete.
Some augmentations such as incorporating the notions of change and motion
into an expanded notion of location- are possible within the space of abstract models. The
key question for augmenting concrete models is
Are there systematic ways of augmenting concrete models that are not reducible
to instances of abstract augmentation mapped on to the concrete sphere?
The answer to this question is Yes. While the concrete world is too rich to
permit a complete classification, there are two important ways, via the notions of are
neighborhood and causal locus, in which concrete representations can be. Both of these
notions are based on concrete metric relations between objects in the world, which are
not reflected in the topological relations encoded in the abstract domain. Furthermore,
both of these notions are ways of augmenting the notion of location in the concrete
domain.
By the neighborhood of an object, I mean the set of objects that are near the
object in some well defined metric space. The metric can vary, and depending on the
metric used, we can get different neighborhoods for the same object. Nevertheless, the
presence of some metric is crucial, because the notion of neighborhood is dependant on
the existence of continuously varying parameters within a metric space. Some important
choices for a metric are given below: Throughout this exposition, I will use the symbol
to denote nearness in some metric space.
(a) The Contact metric. Let A,B,C etc be a collection of objects in three dimensions.
Then two objects A and B are similar, symbolically, A B if and only if A
touches B. Given an object, A, the neighborhood of A, denoted N(A) is defined as
:
N
contact
(A) = (B: a sequence A
1
, A
2
, A
n
such that B A
n
A
n-1
A
1
A)
18
(b) The size metric. Let A,B,C etc be a collection of objects in three dimensions. Let
the diameter Diam(A) of an object, A, be the largest distance between two points
on that object. Then the neighborhood, N
size
(A) = (B: UDiam(B) Diam(A)U U U
for some value of U.
(c) The velocity metric. Let A,B,C etc be a collection of objects in three dimensions
each moving with some velocity V
A
, V
B
, V
C
respectively. Let U be the scalar
product between three dimensional vectors. Then the neighborhood N
velocity
(A) =
(B: UV
B
U V
A
- V
B
U V
B
U U U for some value of U (Intuitively, those objects whose
velocity is close to the velocity of A).
By causal locus, I mean the set of objects that an object can effect physically.
Going back to chapter 3, remember that causation was defined in two forms, one positive
and the other negative, as follows:
(1) Positive Cause. An object causes the state S(B) of another object B, if A U (S(B)
U S(B)), i.e., the presence of A leads to change of state of B from S to S.
(2) Negative Cause. An object A causes the state S(B) of another object B, if (S(B) U
S(B)) U A or counterfactually, UA U (S(B) U S(B)), i.e., the presence of A is
essential to the maintenance of S(B).
In both cases, one can define the causal locus of A, CL(A), as follows:
Definition. The causal locus of A, CL(A) = (B: a sequence A
1
, A
2
, A
n
such
that A causes A
1
causes A
2
causes A
n-1
causes B)
The causal locus is a concrete set of objects because it depends on the actual set of
objects in the world that an object can affect. Note that both the neighborhood of an
object as well as its causal locus offer ways of augmenting concrete global models by
replacing the model of an object by its neighborhood (in some metric) or its causal locus.
Now let us see how we can apply these two means of augmentation to the semantics of
prepositions.
As before, since prepositions are locational, we should expect that figure and
ground models are augmented differently. The fundamentally, the idea is to include as
many details about the location of the figure and as many details about the possible
locations that the ground can generate. In the abstract case, the figure model was
augmented to include dynamic concepts of figure location and the ground model was
augmented to include subtler geometric aspects of its shape.
In the concrete case, in the figure model, we can augment the notion of figure
location by including not only the location of the figure itself with respect to the ground,
but also the location of the neighborhood of the figure with respect to the ground.
Similarly, we can augment the ground model by including (in a coordinate frame with
respect to the ground) not only those locations generated by the geometry of the ground
alone, but also those locations that are part of the causal locus of the ground. While
neighborhoods and causal loci are unique to the concrete domain, augmentation using
neighborhoods and causal loci is consistent with the augmentation procedure in the
abstract domain. This is because change and motion in figure objects is often the result of
causal influences originating in the ground object and these influences are likely to be
transmitted to the figure object via other objects that are part of the figure objects
19
neighborhood (as happens in figures 5.7(a) and 5.7(b). Therefore, if needed, one can
change the meaning postulates for spatial prepositions as follows:
Prepositional Meaning Postulate: Let P(figure, ground) be a spatial preposition. Let
A be its abstract meaning, i.e., A is a node in a partial order generated by an abstract
global model G
abstract
. Let G
concrete
be the concrete global model that G
abstract
maps on to
and let A = A(figure, ground) be the concrete node in G
concrete
corresponding to A (in
G
abstract
). Then a concrete feature, f
concrete
U Meaning(P) if and only if f
concrete
U
A(neighborhood(figure), causal-locus(ground))).
Now, we can see how the prepositional meaning postulate accounts for the
judgements in figures 5.7(a) and (b). In 5.7(a) while the left hand picture satisfies
CONTAIN(N
contact
(black egg), causal-locus(bowl)), the right hand picture does not.
Similarly, in 5.7(b) the left hand picture satisfies CONTACT( N
size
(black book), causal-
locus(table)) while in the right hand picture it does not.
This concludes my presentation of the semantics of spatial prepositions. As I
have shown here, the meaning of prepositions is rather intricate it requires the
understanding of abstract models, concrete models and the mapping between the two, as
well means of augmenting abstract models and concrete models. Once all of these stages
of modeling are available, we can generate a rich set of prepositional meanings one that
is adequate to capture meanings that have been hard, if not impossible, to generate
systematically using other techniques. Nevertheless, the derivational approach offers a
systematic method to understand and generate all of these stages in the semantics of
prepositions suggesting that something category like must be quite central to semantics of
spatial concepts in language.
4.4 A full derivation for ACROSS
So far in this chapter, we have seen how the semantics of prepositions can be
modeled using the idea of a derivation. Now, I show how the notion of derivation can be
used to understand the semantics of individual prepositions in greater detail. I will restrict
my analysis to the semantics of the preposition ACROSS. Let us see how we can
generate the meaning of ACROSS.
Step 1. The abstract, core meaning of ACROSS.
First, I assume that the world is two dimensional and Euclidean. Let us also
assume that the Figure objects are modeled as points moving along smooth curves and
that the Ground objects are modeled as rigid, static regions that partition two dimensional
space. More formally,
(1) SPACE = R
2
.
(2) PATH(Figure) = Smooth curve in R
2
.
(3) The compliment of the Ground, i.e., R
2
Ground is disconnected. For
simplicity, we can assume that the compliment of the Ground has three
components, one Interior Component, I, and two Exterior Components, E
1
and
E
2
(figure 5.8).
20
The product model for the figure and ground define the world that starts our
derivation. Let this world be called W. Then, we can classify the topological states of
W as follows:
Definition: Let I, E
1
and E
2
be the connected components of R
2
Ground as given
above. Then, let the Crossing Number, C(Fig, Gr) = Number of times
PATH(Figure) goes from E
1
to C
2
or from E
2
to E
1
.
Then, the set of distinct topological states of W is classified by C(Fig,Gr), i.e., each
topological state of W has a unique crossing number. Therefore, the abstract meaning of
ACROSS can be defined as follows:
Core Meaning Postulate. ACROSS(Figure, Ground) C(Fig,Gr) =1.
Of course, there are an infinite number of possible configurations that are
compatible with this meaning, some of which are given in path 1-3 in figure 5.9(a) below.
At the same time, the above meaning postulate rules out an infinite number of paths as
well, as illustrated in figure 5.6(b).
21
1
2
3
Figure 8(a)
1
Figure 8(b)
1
Figure 8(c)
Exterior, E
1
Exterior, E
2
Interior, I
Figure 7
Therefore, the abstract meaning for ACROSS does not quite capture all possible
nuances of its meaning. For example, the configuration in figure 5.9(c) above is not a
case where we would normally license the use of ACROSS. In order to understand the
semantics of ACROSS comprehensively, we have to augment the abstract meaning,
which at this stage is purely topological.
Step 2. Augmenting the abstract meaning.
Remember that there are two highly invariant forms of representations topology
and coordinate frame representations. Therefore, coordinate frame representations can be
used to augment topological representations. Let us see how we can do this in the case of
ACROSS.
First, note that in the world model, W, illustrated in figure 5.8, there is a natural
subdivision of the regions of the model. First, there is the bounded component, I, where,
independent of the size of the region enclosed by the ground, we know it is finite. Then,
there are the unbounded components E
1
and E
2
which extend to infinity (in the model).
Therefore, one can define two natural coordinate frames:
(a) CF
int
: This is the coordinate frame based on the interior of the ground, which
is bounded by the boundary of the ground.
(b) CF
ext
: This is the coordinate frame based on the exterior of the ground, which
is unbounded.
Therefore, we can define augmentations of W,namely,
(a) W* CF
int
.
(b) W* CF
ext
.
Which in turn leads to augmented meanings of ACROSS, given as follows:
(c) ACROSS
int
(Figure, Ground) (C(Fig,Gr) =1) (PATH(Figure) CF
int
)
(d) ACROSS
ext
(Figure, Ground) (C(Fig,Gr) =1) (PATH(Figure) CF
ext
)
Already, note that the derivation, while based on the same principles, leads to two
different fleshed out versions of ACROSS. As we start adding more concrete details,
such as the geometry of the path of the figure and the shape of the ground, the meaning of
a preposition becomes more interesting and far more dependent on its context. I argue
that the abstract meaning of a preposition is fleshed out in different ways depending upon
the context of the preposition.
Step 3. Mapping abstract meaning to concrete representations.
Let us see if we can understand the fleshing out procedure in the case of
ACROSS, which we have modeled in terms of connected regions in two-dimensional
space. In step 2, we saw that there are two natural ways to flesh out the meaning of
ACROSS, depending on the use of an external versus internal coordinate frame. In order
22
map the abstract meanings, ACROSS
int
and ACROSS
ext
on to concrete representations,
we need to introduce some metric information in the form of a scale constraint.
Scale Constraint. Let size(G) be the diameter of the ground. Let size(W) be the
diameter of the world, i.e., the extent of the two dimensional space containing
both the figure as well as the ground. Then, assume that size(W) >> size(G).
Given the scale constraint, there are two natural kinds of paths. Let size(PATH(figure))
be the length of the path of the figure. Then, we can define two classes of paths:
(a) size(PATH(figure)) >> size(G)
(b) size(PATH(figure)) size(G)

Now, let us see how these considerations lead to two different fleshed out meanings
of ACROSS, the first of which is a coarse scale meaning, and the other meaning a fine
scale meaning. To take a look at the large scale meaning first, note that if (a) above is
valid, at a coarse scale the shape of the ground object and the detailed relationship
between the figure and the ground can be ignored because their geometry does not show
up at a coarse scale. The only thing that matters is the geometry of the path itself, since
that is the only geometrical detail that is visible at that scale.
Definition. Let C be a curve in 2D and let be a scale parameter where. Let C

be
a smoothed ot version of C, let us say, C

is C convoluted with a Gaussian, G

,of
width (more formally, C

= C* G

). Then define the net curvature, N

,(C) of C
at scale as follows:

Then, we can define the coarse scale meaning of ACROSS as follows:
Coarse Scale Meaning. Let the coarse scale, , satisfy >> size(Ground). Let
be a real number greater than 0.Then, for some ,
ACROSS(Figure, Ground) (C(Fig,Gr) =1)( N

,(PATH(Figure)) ))
The above fleshing out takes care of the coarse scale meaning. What about the fine scale
meaning? The fine scale meaning is a little more complex, since there are more geometric
factors that need to be taken into account. These geometric factors are:
(1) The shape of the ground object.
(2) The curvature and the velocity of the path of the figure object.
How can we incorporate these geometric factors into the meaning of ACROSS? Let
us first note that in determining the meaning of ACROSS, we need to take into account
two factors:
(1) Whether the figure object crossed the ground object or not. In other words, we
have to map the (topological) interior and exterior components of the world with
respect to the ground, I, E
1
and E
2
.on to some metric features of the ground
object. In other words we have to determine the sides of the ground object. Then,
we have to determine whether the path of the figure object went from one side to
23

C C N
2
1
) (
another or not. This determination is primary, for if the figure object did not go
from one side to another, it does not matter how it moved.
(2) If we have determined that the figure has indeed gone from one side to another,
then we can incorporate the manner in which it did so. In this case, we have to see
how the path of the figure is located with respect to the ground as well as the
curvature of the path.
Steps 1 and 2 above are accomplished as follows:
(1) The sides of the ground object are determined by looking at the lattice of
features of the boundary of the ground object. That is to say, they are
determined by the curvature extrema of the boundary. Positive extrema (+
in figure 5.10) demarcate side boundaries while negative extrema
determine part boundaries (- in figure 5.10).
(2) Once the sides have been established and the validity of ACROSS has been
determined, the centroid construction determines the relationship of the path of
the figure to the ground. The centroid construction is given as follows: For any
point, x, at the boundary of the ground, there is a unique point, y, such that a
straight line from x to y divides the area of the ground equally. The path of the
figure should be close to the centroid path. Finally, we should also expect the path
of the figure to be not too curvaceous. The terms close to and curvaceous can e
made precise using the notion of scale (see below).
Combining these analyses, we should get the following model for the meaning of
ACROSS in the fine scale case.
Fine Scale Meaning. Let the fine scale, , satisfy << size(Ground). Let be a
real number greater than 0.Then, for some ,
ACROSS(Figure, Ground) (C(Fig,Gr) =1)( N

,(PATH(Figure)) )
CONNECT(PATH(Figure), SIDE-A(Ground), SIDE-B(Ground)), where the
edges of the sides are curvature extrema of the boundary of the ground.
The fleshing out of the concrete meaning of ACROSS also leads to specific empirical
predictions. First, note that the concrete model is a combination of non-accidental metric
features of the shape. Let us see what this means for two shapes ellipses and squares.
(c) For an ellipse, the non-accidental features are the curvature maxima, therefore
the non-accidental partitioning of the ellipse into sides should be given in
24
+
+
+

Figure 9
terms of the major and minor axes of the ellipse. The major and minor axes
determine the SIDE parameters in the fine scale meaning in the case of
ellipses. Furthermore, the ratio,
Minor
Major
, of the major and minor axes should
determine the acceptability of ACROSS.
(d) For a square, the non-accidental features are the edges and the equality of
sides. Therefore, the SIDE parameters in the fine scale meaning should be
determined by the edges. Also, since the sides are of equal length, the choice
of sides should make no difference in the acceptability of ACROSS.
Step 4. Augmenting the concrete model.
To complete the derivation of the meaning of ACROSS, we need to see how the
concrete meaning can be augmented. The concrete models can be augmented in two
ways:
(1) Imposing a metric coordinate frame on the representation of the geometry.
Coordinate frames can either be observer or object centered. The non-accidental
features represented in a coordinate frame would be the axes (for an object
centered frame) and verticality, horizontalness and alignment with the body of the
observer, for an observer centered coordinate frame. As I mentioned earlier in the
discussion of the experimental results for squares, an observer centered frame
may be implicitly involved in judging the acceptability of ACROSS even though
the default coordinate frame is an object centered one.
(2) Augmenting geometry to include physical structure. For ACROSS, this
can be done by using the idea of a causal path locus. By causal path locus, I mean
the following:
Definition: Let G be the ground object, and let O
1
, O
2
O
n
be objects located in the
interior Int(G) of G. Let be a parameter, where << diameter(Ground). Let N

(O
i
)
be a circle of radius around the object O
i
. Let R(G) be the complement of the circles
in the interior of the ground.
) ( ) (
i
i
O N G G R
U
=
Let F be the figure object. Then the causal path locus, CPL(F, G), of F with respect to
the ground, G, is defined as
CPL(F, G) The set of all paths, P

(F) of the figure such that P

(F) R(G).
Then, we can define the augmented concrete meaning of ACROSS as follows:
Augmented Concrete Meaning. Let the fine scale, , satisfy << size(Ground).
Let be a real number greater than 0.Then, for some ,
ACROSS(Figure, Ground) (C(Fig,Gr) =1)( N

,(PATH(Figure)) )
CONNECT(PATH(Figure), SIDE-A(Ground), SIDE-B(Ground))
(PATH(Figure) CPL(F, G))
The augmented concrete meaning of ACROSS allows us to explain why the path in
figure 5.15 is acceptable.
25
This ends my presentation of the semantics of ACROSS. I have shown how the
derivational account of the meaning of prepositions leads to a detailed analysis of the
meaning of ACROSS, accounting for all shades of meaning, from the core, abstract
meaning based on the world model illustrated in figure 5.8, to the concrete, augmented
meaning illustrated in figure 5.15.
4. Discussion and Summary
So far so good you might say, but what does this have to with conceptual change at the
level of history etc? For that we have to make a couple more detours. First, note that
conceptual change can happen by means of beliefs. I.e., the enaction of a concept may
well involve a belief that concept participates in. Usually, when one asks what is a
concept, we are given an answer of the kind: this concept is so and so which as a
definition is incomplete but as a belief is perfectly fine. And when we revise that belief,
we will also revise our usage of that concept. So beliefs are the link between concepts
and larger level conceptual change. In this sense, my account of concepts is like the
theory-theory of concepts, in that I postulate that a concept is embedded in a network of
beliefs about that concept. Except that that network is embodied and enacted thats the
fundamental difference in my case it is in the nature of that network. The theory
theorist would say that the network is a network of proto-theoretical claims involving the
concept while I would say that for one that theories cannot account for vagueness, and
that theories are a small subset of beliefs as a whole, and beliefs that are not theory like
are as capable of causing conceptual change. For one, beliefs may not be statements
about the nature of the world they may well be commands to act in one way or the
26
Figure 10: John walked across the park
other. If you ask me wheres the cup and I reply (as one usually would) its in the
living room, I am not making a statement about the existence of the cup, but implicitly
about what you have to do get the cup. The enactive approach therefore generalizes the
theory-theory immensely while retaining its explanatory capacity.
Note that conceptual change need not just be historical, it can also be developmental
and in both cases it may be triggered by the change in our beliefs. I want to stress this to
make it clear that there several temporalities involved in conceptual change and that we
need a better understanding of time in cognition in order to account for all of these
temporalities. Thats one of the deep puzzles I wanted to raise at the end of this talk
what are the various dimensions of conceptual change? History and development are two
clear examples. How are they related? Does Ontogeny recapitulate Phylogeny? Second:
If the meanings of concepts are really objects, then where do these objects lie? They
surely are not perceptual objects, so are they just in the mind. No, because they are
shared. So are they inter-subjective objects? But they are also enacted in the external
world. So what kind of objects are they? What is an object anyway? And if these objects
are neither in the mind, nor in society or in the world, where do they come from? I hope
that at the end of this talk, you would be willing to admit with me that Platos problem
that I outlined at the beginning of the talk is wide open and well worth an investment of
our scientific and philosophical efforts.
27